JPH0773275A - Character recognition device - Google Patents

Abstract

PURPOSE:To improve the recognition accuracy of a pattern of large curvature by recognizing mainly the curvature feature value of a contour picture element with a curvature changing point position defined as a feature. CONSTITUTION:A curvature feature value calculating part 10 calculates the complicacy of a character image and the curvature feature value of a contour line. A normalizing part 11 acquires a normalized size of an input pattern. A correlation calculating part 12 calculates the correlation between the curvature feature value of the normalized input pattern and a curvature feature value dictionary 13 and calculates the curvature resemblance. A directive feature value calculating part 14 calculates the directive feature value in a method similar to the conventional one and by the directive differentiation of 4 or 8 directions. A normalizing part 15 normalizes the size of the directive feature value, and a correlation calculating part 16 calculates the correlation between the directive feature value and the character stored in a directive feature value dictionary 17 and to be recognized. Then a general resemblance calculating part 18 multiplying the curvature resemblance and the directive resemblance by a proper weight coefficient and adds both resemblance together to calculate the general resemblance. A sorting part 19 sorts the recognized characters and transmits the character recognizing results.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device for recognizing characters using directional features.

[0002]

2. Description of the Related Art Character recognition methods using an optical character recognition device (OCR) include a method in which a feature amount is obtained by normalizing an original pattern, and a method in which an image resulting from directional differentiation is used as a feature amount. There are various methods such as a method in which the direction / distance from a cell arranged at a fixed position to each stroke is used as a feature amount, and a method in which a stroke structure is analyzed. As one of them, there is a method in which the directionality of the boundary portion between the character pattern and the background is used as the feature amount, and the matching degree between the features is measured by the similarity cos θ obtained from the correlation value of the feature vector space of the two patterns. .

[0003]

This two-dimensional feature correlation method is expressed by using a stroke or a background of a character as a feature amount of a directional differential image in 4 or 8 directions. It has been confirmed that it is effective for printing Kanji in Mincho and Gothic fonts, which are often used. However, even in the case of printed Chinese characters, it is not possible to deal with characters in hiragana or textbooks containing a lot of gentle curve components only with such directional features in four or eight directions, and there is a problem that the recognition accuracy deteriorates. .

Further, even with respect to handwritten characters, there is a limit to the recognition accuracy only with the directional features in a limited direction. In order to solve this, there is also a method of increasing the number of steps in the direction to 16, 32, but since the original images are arranged in a grid pattern, the directional feature quantity is added to the direction beyond 8 directions. There is a problem in that not only the accuracy of itself decreases but also the number of dimensions of the feature amount, the dictionary capacity, the recognition time, etc. increase. Therefore, an object of the present invention is to improve recognition accuracy while suppressing increases in the number of dimensions of feature quantity, dictionary capacity, recognition time, and the like.

[0005]

In order to solve such a problem, according to the present invention, in a character recognition device for recognizing a character by using a directional characteristic amount, characters are extracted from a character image cut out character by character. Size extracting means for extracting the vertical and horizontal sizes of the, a complexity calculating means for calculating the complexity of the character from the line density of the character image, and a directional feature quantity for extracting a two-dimensional directional feature quantity in four or eight directions. Extraction means, smoothing condition determining means for determining the smoothing condition of the contour pixel from the character size and complexity, the curvature of the contour pixel is determined under the determined smoothing condition, and a portion having a large absolute value is a feature amount. A curvature feature amount extraction means for extracting as a straight line, a straightness degree calculation means for calculating a straightness of a stroke forming a character from the variance of the curvature values, and an unknown pattern and a word using the directional feature amount and the curvature feature amount. Obtains the correlation values between the pattern,
It is characterized in that a total similarity calculating means for obtaining a total similarity by multiplying these by a weighting coefficient based on the linearity and adding them is provided, and characters are recognized from the total similarity.

[0006]

In addition to the conventional directional feature amount, the curvature of the boundary component is introduced as a feature amount. The curvature of the boundary component has a large error as it is and cannot be used as a feature amount. Therefore, the complexity of the character is first measured by the line density or the like, and the smoothing coefficient of the curvature is changed according to the size and the complexity of the character. Then, character recognition is performed by integrating the directional characteristics and the curvature characteristics thus obtained. In this case, the weights of the directional feature and the curvature feature are determined by the linearity obtained from the unknown pattern. When the pattern includes many straight lines, the specific gravity of the directional feature is increased. Weigh each and recognize each one.

[0007]

1 is a block diagram showing an embodiment of the present invention. In the figure, 1 is a document input to OCR, 2
Represents an image scanner as an image input device, and 3 represents the entire OCR. A character image input unit 4 controls the image scanner 2 to store the document image in the image memory. 5
Is a line cutout unit that extracts a line in which text is written from a document image. Reference numeral 6 denotes a character cutout unit which cuts out an image of each character.

Reference numeral 7 is a 1-character recognition block, which recognizes the 1-character image cut out in 6. 8 is a portion for removing noise from the character image, 9 is a size for measuring character size and line density from the noise-removed character image.
It is a complexity measuring unit. Reference numeral 10 is a curvature feature amount calculation unit that calculates the curvature feature amount of the contour line from the complexity and the character image, and 11
Is a normalization unit that normalizes this to a prescribed size, 13 is a dictionary that stores the curvature feature amount for each character to be recognized, 12 is the curvature feature of the input pattern normalized by 11, and 13 This is a correlation calculation unit that performs a correlation calculation with the curvature feature amount dictionary to obtain a curvature similarity.

Reference numeral 14 is a directional feature quantity calculating section for calculating a directional feature quantity by directional differentiation in four or eight directions by a method similar to the conventional method, and 15 is a normalizing section for normalizing the size of this feature quantity. , 16 are correlation calculation units for obtaining a correlation value between the directional characteristic amount thus obtained and the directional characteristic amount of the character to be recognized stored in the directional characteristic amount dictionary 17. Reference numeral 18 denotes an overall similarity calculating unit that obtains the overall similarity by multiplying the curvature similarity and the directional similarity by an appropriate weighting factor and adding them. Reference numeral 19 denotes a sorting unit that performs sorting in order of increasing total similarity and outputs a character recognition result, 20 denotes a post-processing unit that performs post-processing such as knowledge processing on the recognition result in character units, and 21 denotes a final processing unit. The recognition results obtained are shown in.

FIG. 2 is a flow chart for explaining the recognition procedure in FIG. First, noise is removed from the input pattern (step S1), then the character size (height, width) is measured at step S2, and the complexity F is measured at step S3. After that, on the one hand, the extraction of the directional feature (step S4), on the other hand, the extraction of the curvature feature of the contour pixel, and the calculation of the linearity αp of the pattern (step S6).
And normalize each (steps S5, S
7) Then, the correlation value Dj of the directional feature amount with each dictionary category is calculated (step S9), and the curvature feature amount correlation value Cj with each dictionary category is calculated on the other hand (step S10).
The total similarity Tj is obtained by weighting these with the linearity αp of the input pattern and adding them (step S11). Sorting is performed in descending order of total similarity (steps S12 to S14), and step S1.
At 5, the recognition result is output.

FIG. 3 is a flow chart for explaining the dictionary creating procedure in FIG. First, noise removal, character size (height, width) measurement, and character complexity F measurement are performed in order from the first (i = 1) pattern in the category (steps S1 to 4). After that, 2 in parallel
Calculate two feature quantities. On the other hand, the directional feature quantity is extracted and the size is normalized as in the conventional case, and the obtained feature quantities di _{1 to} di ₈ are added to the dictionary area (steps S5 to 7). On the other hand, a smoothing section that is proportional to the character size and inversely proportional to the complexity is set, and the curvature image of the contour pixel is obtained (step S8). Also for this, the size is normalized in the same manner, and the obtained feature amount c
i ₊ ~ci _- a dictionary respectively C _+, C _- the Komu added (Step S9 and S10). This is performed for all patterns in the category and then normalized, and dictionary creation for one category is completed (steps S11 to 13).

A detailed description will be given below. FIG. 4 shows an example of a character image. This is an example of the kanji "Oto" in the textbook. This character consists of a straight line and a curved line, but the way the curve bends differs depending on the typeface, and is often misread as the hiragana “te”. The complexity F of this character is
For example, it can be obtained from a function that counts the rising of a line as shown in the following equations 1 to 5.

[0013]

[Equation 1]

[Equation 2]

[Equation 3]

[Equation 4]

[Equation 5]

The complexity obtained from the character image shown in FIG.
It looks like this: Fx = 1.38 Fy = 2.67 F = 2.04

A method of extracting directional features (strokes) in eight directions by the conventional method from the character image shown in FIG. 4 will be described. FIG. 5 is an explanatory view for explaining 8 directions, FIG. 6 is a 3 × 3 local memory block diagram for extracting directional features, and FIG. 7 is an example of a directional differential operator for extracting directional features. 8A and 8B are examples of differential images corresponding to FIG. 7 (a) and 7 (b) are examples of the well-known Sobel operator, other operators capable of extracting the direction can be substituted.

First, in the local area of FIG. 6, FIG.
By multiplying the Sobel operator shown in (b), differential values ∇x (x, y) and ∇y (x,
y) is calculated. An example of the differential image with respect to FIG. 4 is shown in FIG.
Furthermore, using this, the direction θ of the boundary component
(X, y), that is, the angle is calculated.

[Equation 6]

[Equation 7]

[Equation 8] Boundary component direction θ (x, y) obtained as described above
FIG. 9 shows the result of arranging B in the order of the boundary components.

The curvature is the direction θ at a point I (x, y).
It can be expressed as a variation amount (angle variation amount) of (x, y). Here, the curvature K (x, y) is calculated by the following equation 9. However, (xi, yi) in the equation 9 indicates the coordinates of the outline point sequence, and indicates that the boundary components are arranged in the order of i. Further, FIG. 10 shows the curvature K (x, y) obtained by the following Expression 9 for the image of FIG. In FIG. 10, the smoothing section is set to w = 4, but the smoothing coefficient w is almost proportional to the size of the previously obtained character and inversely proportional to the complexity of the character.

[Equation 9]

This curvature value is normalized, and only the pixels whose absolute value is a certain threshold value (here, π / 2) or more are +
Surfaces (curvature> 0) and negative surfaces (curvature <0), which are returned to the image shape, are shown in FIGS. 11 (a) and 11 (b), which are referred to as curvature feature quantities. The point where this curvature feature amount appears is a point where the change in the direction of the contour is large, the + surface of (a) is a convex surface, and the − surface of (b) is a concave surface. The painted part is the apex part with a large curvature. Further, the arrow (a) indicates the tracking direction of the image.

[0019] Here, ₊ normalized directional feature vectors of d ₁ to d ₈ of unknown pattern, like the normalized curvature feature vectors c, c _- a, and directional feature vector of the dictionary j Are dj ₁ to dj ₈ , and the normalized curvature feature amount vectors are cj ₊ and cj ₋ , the directional feature correlation value Dj of the unknown pattern and the dictionary, and the curvature feature correlation value Cj are It is obtained by the equations 11 and 12. Nk represents the number of dimensions of the feature amount vector Dk. Further, a · b represents the inner product of the vector a and the vector b, and the norm is represented by enclosing the code in double lines.

[0020]

[Equation 10]

[Equation 11]

[Equation 12]

Since the linearity αp proportional to the variance of the curve can be obtained from the curvature feature vector, the directional characteristic correlation value Dj and the curvature characteristic correlation value Cj are multiplied by the linearity αp to obtain the total correlation value Tj as follows. It is obtained by the equation 13.

[Equation 13]

[0022]

According to the present invention, the curvature feature amount of the contour pixel is introduced in addition to the conventional directional feature amount, and a pattern including a large number of straight lines increases the specific gravity of the conventional directional feature amount, and a curved line is formed. For patterns that contain many patterns, the gravity of directional features is reduced, and the curvature change point position (vertices, etc.) is mainly based on curvature features.
Since it is recognized as a feature, there is an advantage that a pattern with many curves can be recognized with high accuracy.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a flowchart for explaining a recognition procedure in FIG.

FIG. 3 is a flowchart for explaining a dictionary creation procedure in FIG.

FIG. 4 is an explanatory diagram showing an example of a character image.

FIG. 5 is an explanatory diagram for explaining the definition of a direction.

FIG. 6 is an explanatory diagram illustrating a local memory.

FIG. 7 is an explanatory diagram illustrating an example of a differential operator.

FIG. 8 is an explanatory diagram for explaining an example of a differential image with respect to FIG. 4.

FIG. 9 is an explanatory diagram illustrating a direction (angle) of a contour image.

FIG. 10 is an explanatory diagram for explaining the curvature of a contour image.

FIG. 11 is an explanatory diagram for explaining a curvature feature amount.

[Explanation of symbols]

1 ... document, 2 ... image scanner, 3 ... character recognition device,
4 ... Character image input section, 5 ... Line cutout section, 6 ... Character cutout section, 7 ... 1 character recognition block, 8 ... Noise removal section, 9 ...
Size / complexity measurement unit, 10 ... Curvature feature amount calculation unit, 1
1, 15 ... Normalization unit, 12, 16 ... Correlation calculation unit, 13 ...
Curvature feature amount dictionary, 14 ... Directional feature amount calculation unit, 18 ... Total similarity calculation unit, 19 ... Sorting unit, 20 ... Post-processing unit, 21 ... Recognition result.

Claims (1)

[Claims] 1. A character recognizing device for recognizing a character using a directional characteristic amount, and a character size extracting means for extracting a vertical and horizontal size of a character from a character image cut out in units of one character, and a line of the character image. A complexity calculating means for calculating the complexity of a character from the density, a directional characteristic quantity extracting means for extracting a two-dimensional directional feature quantity in 4 or 8 directions, and a smoothing condition for contour pixels from the character size and complexity. The smoothing condition determining means for determining, the curvature feature amount extracting means for obtaining the curvature of the contour pixel under the determined smoothing condition, and extracting the portion having a large absolute value as the feature amount, the character from the variance of the curvature values. And a linearity calculation means for calculating the linearity of the strokes, and the correlation values between the unknown pattern and the dictionary pattern using the directional feature quantity and the curvature feature quantity, and the linearity That the overall similarity calculating means for calculating an overall similarity by weighting factors and adding the multiplied respectively, the provided character recognition apparatus characterized by performing the recognition of characters from the overall similarity.